Every team with monitoring has the same two habits. First, someone gets paged during a planned deploy and mutters. Second, someone turns alerts off "for the deploy" and forgets to turn them back on. The second habit is how a real outage runs for six hours unnoticed on a Friday night.
The right tool is a maintenance window: alerts are silenced for a defined period, and — this is the important part — monitoring resumes the instant it ends, so a deploy that leaves something broken still pages you.
What a good window does
- Pauses checks for the monitors it covers, so no downtime is recorded for work you chose to do.
- Has an end. A window with no end is just alerts turned off.
- Scopes to what you're touching — deploying the API shouldn't silence the marketing site.
- Doesn't forgive. When it closes, the normal failure-confirmation rule applies immediately. If the service is still down, that is a real incident.
Open it from the pipeline
Doing this by hand is how it gets forgotten. The reliable pattern is to have the deploy itself open the window, sized to a bit longer than the deploy takes. With UpConduit's API that is one request at the start of the job:
curl -X POST https://cj0dxr3cke.execute-api.eu-west-2.amazonaws.com/v1/maintenance \
-H "Authorization: Bearer $UC_API_KEY" -H "Content-Type: application/json" \
-d "{\"name\":\"Deploy $GIT_SHA\",\"mode\":\"once\",\"start\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\",\"durationMin\":20}"
Or, with the command-line tool: uc maint open "Deploy $GIT_SHA" 20. Twenty minutes later the window closes on its own; nobody has to remember anything. If your deploy finishes early, close it early with uc maint close <id> so monitoring resumes sooner.
Recurring work
For scheduled jobs — the Sunday 02:00 reboot, the weekly database maintenance — set a weekly window once, in your own timezone, and it handles daylight-saving changes without you. Scope it to the monitors that actually go quiet.
What not to do
Don't lengthen the failure threshold to "handle deploys". That makes every real outage take longer to detect, all the time, to solve a problem that happens twice a week for twenty minutes. Windows are precise; thresholds are blunt.
Maintenance windows, the API and the CLI are all on UpConduit's free plan. The API docs have the full recipe.